KR100251000B1 - Speech recognition telephone information service system - Google Patents

Abstract

PURPOSE: A system for servicing voice perceiving telephone information is provided to change a fixed connection structure between a DSP(Digital Signal Processor)0 and a DSP1 as a dynamic connection structure. CONSTITUTION: A system performs an A/D conversion for an inputted voice, and emphasizes a high-pass component of the voice. The system divides the voice as frames of a 20msec length. The system overlaps the frames each 10 msec to prevent an information loss of frame boundary surfaces. The system extracts a cepstral coefficient and various coefficients displaying voice characteristics. The system refers to a vector quantization code book each kind, and obtains an index of values which are mostly similar to vector values. The system transmits the index from a DSP(Digital Signal Processor)0 to a DSP1. The system uses results of previous frames while performing a voice characteristic extract and a vector quantization for one frame, and performs a Viterbi search.

Description

Voice recognition telephone information service system

The present invention relates to a voice recognition telephone information service system in which a fixed connection structure between DSP0 and DPS1 is converted into a dynamic connection structure in a digital signal processor (DSP) module in a voice recognition system.

The voice recognition telephone information service system is a system that allows the general public to request certain information from the system by voice, not DTMF (Dual Tone Multi Frequency) through the telephone network, and provides related information by voice. In the telephone voice information system, the user must memorize the relevant service code or know the menu to receive the desired service. For example, in the case of securities information guide, the user needs to know the code number corresponding to the desired company name and input it, but using the voice recognition function, only the company name needs to be entered. In addition, in the case of natural sound guidance, if you want to hear the magpie from among the birds, you must select the bird sounds menu and select the magpie, but if you have voice recognition, you only need to enter the magpie.

There are several ways to implement speech recognition. The most popular method is HMM (Hidden Markov Model), which maximizes the expected state of the corrected state by selecting the maximum frequency state at every moment among the probabilistic statistical methods. The reason is that it is robust against vocabulary, users, pattern comparison algorithms and word-determination logic, and performs well in all areas.

Telephone voice information system that does not recognize the voice is inconvenient to enter the service code or to know the menu name to find the information you want to know. You can think of a system that adds voice recognition, but the system still has a high cost of hardware, making it an obstacle to commercialization.

In order to solve the above disadvantages, an object of the present invention is to provide a voice recognition telephone information service system with a lower cost of hardware.

1 is a block diagram showing a speech recognition process by pattern recognition.

2 is a diagram of a conventional system structure.

3 is a table showing the DSP pipeline.

4 is a modified system structure diagram.

This system is implemented using a system with a voice recognition function, the process is shown in FIG. However, in the case of using the HMM method, since a large amount of calculation is required, one processor cannot cover all of them in order to provide services in real time, and is usually configured using two processors. The system uses a processor with a DSP chip, but other processors can be used if there is a better processor.

2 shows the configuration of a system equipped with a DSP module. Looking at the functions performed by a DSP module in a conventional independent word isolated speech recognition system, a guide broadcast capable of processing a voice input even during a guide broadcast is shown. DSP0 is responsible for elimination, end point detection to automatically find speech sections, and analysis of found speech sections to find code numbers in vector quantization codebooks held in the system.

In addition, Viterbi search, a technique for finding a sequence of states most suitable for a given input, is performed by DSIP1. DSP chips are designed for use in signal processing and speech recognition, making real-time processing possible even with many mathematical calculations. However, the price of chips and memory used here is still high, which makes it an obstacle to commercialization.

Here, the speech recognition process by pattern recognition will be described in more detail as follows.

First, all signals coming through the telephone network are sampled at 8 kHz and sent to DSP0, where the beginning of the voice is found from the input data. Referring to Figure 1 describes the speech recognition process as follows. Once it is determined that a voice has been received, it is subjected to a pre-emphasis process that specifically emphasizes the higher end of the baseband signal frequency to improve the SN ratio, which a / d converts the voice and emphasizes the high frequency components of the voice. The input voice is then divided into frames of 20 msec length in the window frame, and the frames are overlapped by 10 msec to prevent information loss at the frame boundary. That is, the previous 10msec data and the newly input data 10sec are made into one frame, and a parameter representing frequency characteristics is obtained for the voice included in each frame. The parameter is transformed after linear predicitive coding (LPC) analysis to obtain a cepstral coefficient. Here, linear predictive analysis is a method of linearly manipulating sample values of a normal probability process to obtain predicted values and prediction errors, obtaining spectral decompositions from them, and analyzing the frequency of speech. The coefficient of time. In this process, we extract the coefficients that represent the speech features, including the Sepstral coefficients. These coefficients are obtained by referring to the vector quantization codebook prepared in advance for each type, and obtain an index corresponding to the values most similar to the vector values in the codebook. DSP0 then passes this index to DSP1. At this time, the DSP1 selects the most similar vocabulary among the recognized word models already stored in the database through the Viterbi search based on the HMM. In addition, Viterbi search can be performed using the results of previous frames while extracting speech features and vector quantization for one frame.

As a result, a pipeline is formed between the DSP chips and is shown in Table 3. Where in Viterbi search 1,2,… , n are not independent of each other but they all come together to form a search. The reason why the right side of the Viterbi search section is shown in dotted line in Table 3 is that the search time is not constant. After the word is recognized, the user is informed of the necessary information. During this announcement, the DSP0 is continuously executed to detect a new voice input of the user. On the other hand, in DSP1, since no work is performed during announcement, DPS1 is idle with idle time.

Therefore, if the channel increases, it causes a waste of resources. In particular, DSP1 requires a lot of memory to store probability values for the phoneme model constituting the recognition word and a phoneme table for words.

The present invention dynamically changes the connection between existing DSP chips to eliminate this inefficiency. In other words, the number of DSP0 and DSP1 is divided by the appropriate ratio according to the amount of work to be performed, and the result of DSP0 is allocated to the no-working DSP1. To this end, it is necessary to change the fixed connection structure between DSP0 and DSP1 into a dynamic connection structure. Here, various methods such as shared memory commonly used by a plurality of processors can be used.

In the speech recognition system using the pipeline processing method, the connection player has four connections and the ratio of DSP0 and DSP1 is N: 1, for example, 2: 1, is shown in FIG. Considering an example of a scenario in which the system is operating, referring to Figure 4, when one DSP0 is connected to one channel and starts to recognize the user's voice, DSP0 notifies the processor managing the system's resources. . The processor then delivers the received data to the available DSP1. When the recognition result comes out, it finds the relevant information and outputs it by voice. In the case of audio output, DSP1 is stopped as described above, so data from other DSP0 can be processed. As a result, the voice recognition system using the pipeline processing method of the present invention can determine how many DSP0s can be mapped to DSP1, and adjust the ratio of DSP0 and DSP1 as the channel is expanded. Can be evenly distributed.

In the above description, DSP0 can directly feed data to the DSP1 through a channel, but data can also be transmitted and received through the common memory shown in FIG. When sending and receiving data, which method is used depends on the protocol and a protocol that takes into account the system's specifications should be used. On the other hand, when connecting between ESP chips, if there is a dynamic connection structure other than the bus structure, the bus structure does not need to be used. Even if other structures are used, the connection structure must be dynamic to evenly allocate load.

When the system is running, the DSP1 at rest can process data from other DSP0s as described earlier when it is being spoken, resulting in up to several DSP0s mapped to one DSP1. . In addition, as the channel expands, the load on the system can be evenly distributed by adjusting the ratio of DSP0 and DSP1. This makes the system efficient and reduces the cost of the system.

Claims (6)

A speech recognition system using a pipeline processing method, the speech recognition system comprising: a pre-emphasis step of a / d converting an input voice to emphasize a high frequency component of the voice; Dividing the input voice into a frame of 20 msec length in the window frame; Overlapping the frames by 10 msec to prevent information loss on the frame boundary; Extracting several coefficients representing speech features, including septal coefficients; A step of obtaining the indices corresponding to values most similar to the vector values in the codebook by referring to the vector quantization codebook prepared in advance for each type; DSP0 then passes this index to DSP1; And performing a Viterbi search using the results of the previous frame during speech feature extraction and vector quantization for one frame. 2. The speech recognition system of claim 1, comprising separating the modules performing the feature extraction and the vector quantization and the modules performing the Viterbi search to dynamically connect several channels. The speech recognition system of claim 1, comprising the step of sharing phonemic table information necessary for performing a Viterbi search. 2. The speech recognition system of claim 1, comprising performing feature extraction and vector quantization and a dynamic linking scheme that performs Viterbi search. The speech recognition system of claim 1, further comprising a speech recognition method that uses fewer modules in the Viterbi search step than the number of modules in the feature extraction and vector quantization. The speech processing method according to claim 1, wherein in the step of performing feature extraction and vector quantization, the pipeline processing method for performing the search in the step of performing the Viterbi search using the intermediate result before the end of the vector quantization is used. Recognition system.

Images